Page 59 - Deep Learning

42 Introduction

at some later time t by saying “Wellington,” we infer that sometime between
2
t and t he learned that Wellington is the capital. The situation is similar with
2
1
respect to skills. If a person cannot drive a car at time t , but turns out to be
1
a skilled driver at time t , we infer that he learned how to drive a car in the
2
meantime. Evidence for learning is essentially comparative, consisting of dif-
ferences between two or more behaviors (actions, utterances) occurring some
time apart. But the change in behavior is not in and of itself learning.
If behavior is generated by processes operating on knowledge representa-
tions, then a change in action or discourse must have been preceded by some
change in the underlying knowledge. It is the act of revising the representation
of New Zealand by replacing the capital is Auckland with the capital is Wellington
that constitutes learning. The change in behavior – the different answer given
to the question what is the name of the capital? – is an overt expression of that
internal change. The distinction between a change in knowledge and its overt
expression is necessary because the person might learn the correct name of the
New Zealand capital but live the rest of his life without ever using that piece of
knowledge; nobody ever asks him, New Zealand never comes up in conversa-
tion and so on. The situation is similar with respect to a skill such as driving a
car. The change of interest is a change in the person’s representation of the skill;
any observable change in his driving behavior is a consequence of that inter-
nal change. The relevant unit changes are changes in knowledge, not changes in
behavior.

The Repertoire of Learning Mechanisms

Sciences differ in an interesting way with respect to their repertoires of unit
changes. Some fields of research attribute all their phenomena to a single type of
42
change. Adherents of the mechanical world view in the 17th and 18th centuries
tried to explain all of physics through the motions of physical bodies and the
forces they exert on each other. Chemists explain all chemical reactions in terms
of the rearrangement of atoms through the breaking and forming of atomic
bonds. Earth scientists, in contrast, draw upon a rich and varied repertoire of
change mechanisms: glaciation in response to astronomical cycles, plate tecton-
ics, erosion caused by water freezing in cracks, the actions of wind and water on
soil and sand and so on. Whether a science will turn out to need a sparse or a
43
rich repertoire of change mechanisms cannot be known ahead of investigation.
Tradition has handed down a long list of suggestions about the basic pro-
cesses of knowledge change. I refer to them as learning mechanisms. Perhaps
association, the idea that knowledge changes by the creation of a link between